The cool thing is we can now stop relying solely on human imagination for ideating on futures funnels, and leverage GPT-4 to help us do it, and get a much more

The cool thing is we can now stop relying solely on human imagination for ideating on futures funnels, and leverage GPT-4 to help us do it, and get a much more realistic sense of what types of things an AI system might conceivably attempt given we can probe these models for what is in them. Granted that's not perfect because GPT-N+1 will have emergent capabilities not present in GPT-N and so it's not always going to be a realistic assessment, but it's an ever closer approximation and it gives much more information than relying solely on human imagination. I've been doing this a bit lately and despite the fact that I know it's only predicting the next token, I'm still slightly unnerved as I watch a black box computer system slowly type out novel and plausible ways to kill me.